How to compare API prices with prompt caching

Before you start

• Use the current selected-model rates per million tokens; cache support and tariffs differ by model.
• Take token counts from Logs and keep ordinary input, reads, writes and output mutually exclusive. Check whether a client's total input already includes cached tokens.

Steps

1. Calculate the bill

   Cost = (ordinary input tokens × input rate + cache-read tokens × read rate + cache-write tokens × write rate + output tokens × output rate) / 1,000,000. Do not count cached tokens twice.

2. Compare the same workload

   The no-cache baseline for the 40/40/20 scenario is 0.8 × input rate + 0.2 × output rate. Both represent 1M total tokens and the same output share; the cache scenario excludes initial cache writes.

3. Use your actual mix

   Include initial cache writes, misses and later reads from a representative set of requests. Effective cost per million total tokens = total cost / total tokens × 1,000,000. Compare providers using the same model, token mix and billing categories.

4. Check when cache writes pay off

   For a repeated prefix of P tokens, one cache write at rate W followed by N reads at rate R costs P × (W + N × R) / 1M. Sending the prefix as ordinary input N + 1 times costs P × (N + 1) × I / 1M. The write is cheaper only when N × (I − R) > W − I. Use the selected model's rates; expiry, changed prefixes and other prompt tokens alter the result.

Current ApiClo model rates and example cost

USD per 1 million tokens. The example uses 40% ordinary input, 40% cache reads and 20% output; the baseline uses 80% ordinary input and 20% output. Initial cache writes are extra. These are calculated examples, not observed customer averages.

Scroll the table horizontally for all rates

Current tariff details

Check the result

Use the current model-by-model rates and scenario costs in the table below, then compare your estimate with Logs. A lower cache-read rate saves money only on tokens actually served from cache.

Troubleshooting

• Treating a cached-input price as the price of all tokens understates output and uncached-input cost.
• Omitting cache writes can overstate savings on short sessions; writes can cost more than ordinary input.
• An assumed cache-hit share is not a guaranteed discount or an observed average.
• Do not describe a provider as universally cheapest without a current comparison of the same workload and capabilities.

Compatibility and billing

The scenario excludes cache writes, hosting, client subscriptions and other non-token charges. It is not a guarantee of cache hits. ApiClo's current rates and actual usage determine the charge; the formula supports comparison without promising the cheapest provider for every task.

Sources

• Anthropic prompt caching
• OpenAI prompt caching